Raw LLM Responses

Did you know that current publicly available chatbots are better than 94% of expert virologists at troubleshooting wet lab steps relevant to creating a bioweapon? Did you know that AI systems are now better than almost all (and in some cases all) humans at the world's most prestigious math and coding competitions, when given the same constraints? That an AI system is now the world's best hacker, by number of vulnerabilities found? That an ensemble of AI models in a scaffold outperform a panel of five human physicians at clinical diagnosis by 4x, at less time and testing costs? There is a lot of stuff they cannot do, and real world utility lags the benchmarks. But real world utility is also catching up. AI capabilities have yet to slow down. There was a useful but somewhat flawed study that exposed that self-reported speedup from AI coding is not reliable, and the developers were slowed down by as much as they thought they were being sped up. It was also a worst case scenario for the AI in doing the study the way they did, with large complex code bases and experienced devs who have been working on them for years, and who hadn't used that specific AI tool before. Meanwhile, many companies are actually measuring real productivity boosts from integrating AI coding assistants into their workflows. I am saying this not as an AI fanboy, but as someone who has deeply followed the technology and actually read papers and thought about their implications for years. If the bubble pops and AI is declared defunct as a technology tomorrow, I will party in the streets. But I volunteer with PauseAI because I can't count on that happening. I have dived very deep into the arguments presented by Eliezer Yudkowsky (long before he wrote the book), and he seems to me to be obviously correct in his core arguments, many of which have been empirically validated even at this stage.